Double SoC approach, integrated 5G modem

Over the past three years, Huawei has announced its next-generation SoC at the IFA Technology Fair in Berlin. On each occasion, the company promotes its equipment using the latest process technologies, the latest design concepts and its latest connectivity options. The announced Kirin flagship processor is found in all major Huawei and Honor smartphones for the next year, and the Kirin 990 family announced today is no different. With the launch of Mate 30 on September 19th^th, Huawei has raised the lid of its new flagship chipset, with some twists.

Double SoC approach: Kirin 990 (4G) and Kirin 990 5G

As we enter the 5G era, we have a two-pronged market. On one side, some regions are not ready for 5G, and consumers in this country do not want to pay the extra money, nor the potential power or trade-offs in a device in order to take over the 5G. Other regions are surfing the 5G wave and are at the forefront, and could therefore pay the premium. Rather than offering a unique solution to both markets, Huawei is sharing its strategy for the first time with two versions of the Kirin 990.

These versions will officially be known as Kirin 990 and Kirin 990 5G. The (4G) I put here is simply to add a differentiator to differentiate them. The two Kirin chipsets and a standard base level are pretty much the same. Same basic configuration, same camera support, same memory, same storage. However, in a few key areas beyond the modem, there are differences, such as the performance of the NPU and the main frequencies. We will come back to this in a moment. But it should be noted that the Kirin 990 5G version is a vision of the future.

The Kirin 990 5G: the SoC of the future

We are constantly talking about 5G, because that is where infrastructure and mobile investments go a lot. Back at the Mobile World Congress in February, we covered all the companies that had announced their own 5G modem, a chip added to a device in order to activate the 5G. This usually meant that we had a standard processor chip with 4G, and then an additional 5G support chip. Ultimately, to achieve the best performance, the 5G chip must be integrated on the same silicon, which allows a better efficiency in 5G mode in exchange for the complexity of the area and design.

True to their reality, Huawei (and its design arm, HiSilicon) are the first to do so for the smartphone market.

The Kirin 990 5G is a true unified design, supporting 5G networks below 6 GHz on SA and NSA architectures. In order to control chip size, Huawei uses TSV's latest TSV + manufacturing process with EUV, which reduces the size of chips for the types of devices in which this chip will be used.

To date, neither Qualcomm, nor Samsung, nor Apple, have a unified flagship chip design close to commercialization. We expect that they will publish the material during the update from generation to generation, but so far, Huawei is the first to announce it.

Thus, despite the fact that a single SoC smartphone can support 4G and 5G without additional hardware, Huawei nevertheless felt it prudent to produce a separate chip without the 5G, all the more so as it was not possible to use it. 5G adoption is still ongoing in the world and that there are still a few years markets on which Huawei is competing. This also allows Huawei to split some of its features and save the best for 5G hardware.

The Kirin 990 series: Details

As mentioned earlier, one of the key elements of the Kirin 990 5G is its use of the 7FF + of TSMC with EUV, which allows the chip to have a smaller matrix size. We are told that the chip is more than 100 mm2, an increase of 74.13 mm2 on the Kirin 980 (TSMC 7 nm) and 96.72 mm2 on the Kirin 970 (TSMC 10 nm), which probably makes Huawei's largest SoC for smartphone to date. This is compared to the Kirin 990 4G version, which measures about 90 mm2, but is based on the same 7 nm process as the Kirin 980, which makes it a bit bigger. The number of transistors for both chips puts the 990 5G at 10.3 billion, while the 990 4G represents about 8 billion.

CPU

The configuration of the cores on both SoCs is identical: two high-frequency A76 cores, two medium-frequency A76 cores, and four more efficient A55 cores. These are divided into their own areas of power and frequency, which allows for greater flexibility depending on the workload. However, the frequencies of the 990 5G and 990 4G will both have slightly different frequencies, depending on the differences between the 7 and 7+ processes.

Huawei Kirin 990 Family
AnandTech	Kirin 990 5G	Kirin 990 (4G)	Kirin 980	Kirin 970
CPU	2xA76 to 2.86G 2xA76 to 2.36G 4xA55 to 1.95G	2xA76 to 2.86G 2xA76 @ 2.09G 4xA55 @ 1.86G	4xA76 @ 2.60G 4xA55 @ 1.80G	4xA73 @ 2.36G 4xA53 to 1.80G
GPU	G76MP16 @ 700M	G76MP16 @ 700M	G76MP10 @ 750M	G72MP12 @ 850M
NPU	2 + 1 Da Vinci	1 + 1 Da Vinci	2 Cambricon	1 Cambricon
Modem	Balong 5G	4G	4G	4G
DRACHMA	LPDDR4-4266 + LLC	LPDDR4-4266 + LLC	LPDDR4X-4266	LPDDR4X-3733
Size of the matrix	> 100 mm2	~ 90 mm2	74.13 mm2	96.72 mm2
transistors	10.3b	~ 8.0b	6.9b	5.5b

For caches, the four A76 cores have 512 KB of L2, while the A55 cores have a capacity of 128 KB each.

Technically, Huawei calls "A76-based" A76 cores because some improvements have been made to the core of the cache system to improve memory latency. Huawei would dictate nothing more than to say that its "SmartCache" implementation, which helps the graphics processor, also helps the processor and the NPU. We believe that it is essentially a higher-level cache above the DynamiQ DSU, similar to the implementations of Qualcomm and Samsung.

A side note here: we expected Huawei to launch the new Kirin with Arm's latest A77 core, as announced earlier this year. As a priority partner partner of Arm, the company's technical team explained two things to us: firstly, the main decisions were made nearly two years ago for this chip, but apart from that, they did not see not the expected frequency of the A77 at TSMC. Process 7 nm.

Huawei said that even though the A77 reached higher peak performance, the energy efficiency of the A77 and A76 at 7 nm was virtually identical, but thanks to a better experience of the A76 to 7 nm, they were able to push the frequencies of the heart much higher. It was mentioned that other companies offering the announced A77 products only reached 2.2 GHz with similar process technologies in other manufacturing plants. It was stated that the A77 would probably come on a future product, most likely when 5 nm would become more widely available.

Regarding LPDDR5 support, we've been told that LPDDR5 technology is still an expensive technology, and Huawei is exploring this possibility for future products.

Graphic

For graphics, the Kirin 990s will both have a 16-core Mali-G76 implementation, compared to a 10-core Mali-G76 in the Kirin 980. This is partly because of the increased size of the matrix. : Huawei believes a lower voltage, a lower frequency but a wider GPU will offer a better chip as a whole.

GPU performance has increased from a 750 MHz 10 core design to a 700 MHz 16 core design.

NPU

In addition to the implementation of the 5G modem, the biggest change on the Kirin 990 will be the Neurological Treatment Unit (NPU). As a company, we will not see much Huawei promote this change because it will ultimately be transparent to consumers, but for the technical side, it is a big step forward.

In Kirin 970 and Kirin 980 hardware, Huawei has sublicensed Cambricon Technologies' machine learning machine design license from a university research project in China. Huawei eventually invested in the company, although the hardware license is not exclusive, but Huawei had access to advanced design and customization options. With the Kirin 990, this partnership with Cambricon disappears and the company implements its Da Vinci internal architecture.

We covered Da Vinci architecture at Hot Chips a few weeks ago, where the company unveiled a number of technical details behind the design. Huawei announced that this architecture would end up in everything from additional 300 W server cards to integrated devices the size of a credit card. Huawei's first smartphone chip with a Da Vinci-based NPU was the Kirin 810, but is now focusing on the flagship SoC of the 2019/2020 generation.

What does Da Vinci bring exactly? Two elements, both important for the application of machine learning algorithms.

First, the Da Vinci "big" cores support the quantization of the INT8 and FP16 networks. In the Kirin 980 with the Cambricon design, the double NPU was split, the two cores supporting the FP16, but only one supported the INT8 for technical reasons. This restriction disappears and all the major Da Vinci kernels support both. Quantification support becomes important to provide faster and less powerful solutions to ML inference problems.

The second change is the addition of a new "Tiny Core" nuclear power plant. The 4G and 5G models will have one. This is a smaller version of Da Vinci's energy-efficient architecture (Huawei cites 24-times efficiency) connected via the AXI bus. The performance of the Tiny Core is naturally inferior, but it is a place conducive to a non-critical ML or low polling, such as voice reactivation or charging characteristics. It can even process individual photos, but is not fast enough to reproduce the motives of a live video. For that, you need big kernels.

A key feature of Da Vinci to note is that Huawei said it has already optimized the software stack for 90% of the most popular computer-based neural networks on the market. One of the advantages of the Da Vinci design compared to that of Cambricon is that Da Vinci is fully compliant with the NNAPI standard, while the older version combined a set of applicable features.

This is the change in the NPU, but there is also a difference between the 990 5G and 990 4G. In addition to the modem, the graphics processor and the manufacturing process, one of the contributions to the matrix size difference is that the 990 5G has twice the number of NPU cores. The 990 5G will have two "big" U NPU cores, supporting two ML processes simultaneously, as well as one NPU Tiny Core. In comparison, the 990 4G will have only one "big" U NPU core, plus the Tiny Core.

This means that some features of Kirin 990 5G devices may not be possible on Kirin 990 4G devices. It will be interesting to see how Huawei, as a company, manages this messaging, especially if it ends up offering its flagship device in 4G and 5G versions only.

The Balong modem

In addition to being the first 5G integrated smartphone design, Huawei finally did not give much details about the new 5G modem, nor about the 4G design updates. It has been mentioned that Kirin 990 5G is the first full-bandwidth modem supporting both NSA and SA architectures (although the Exynos Modem 5100 technically holds this title).

They cited modem top speeds of up to 2.3 Gbps and 1.25 Gbps, with additional ML-based beamforming technology that allows for higher speeds when traveling at high speeds. The design will also allow simultaneous connection to 5G and 4G, for weak signal areas. We confirmed that the company still uses Tensilica DSPs. The technical team stated that, despite international concerns, Tensilica's license was still valid.

Huawei's allegations about the performance of Kirin 990 5G

As expected, Huawei wanted to present the performance of the largest SoC compared to current competitors in the market. The 4G model did not appear in many of the graphs that were presented to us.

The overall figures indicated a + 9% increase in single-threaded performance from Kirin 980 to Kirin 990 5G, mainly due to the higher frequency. The overall performance of multithreads was up 10%. However, energy efficiency has increased 35% over average A76 cores compared to last year, and Huawei expects that most non-performance-related workloads will be performed on these central hearts. (To be complete, Huawei says that high-performance cores are 12% more efficient than the previous generation and the smaller ones are 15% more efficient.)

On the GPU side, we expect an increase in performance. However, Huawei refrained from quoting figures within a reasonable period of time so we could note them in our briefing. We were able to get details on how the 'Smart Cache' improves performance: in this case, the Kirin 990 (both versions) sees a 15% reduction in GPU-DDR bandwidth and a 12% reduction in DDR energy consumption (as it uses less in the same workload).

The overall performance is an AI, although the numbers here are going to be split between the 990 4G with a basic 1 + 1 NPU design and the 990 5G with a basic 2 + 1 NPU design. Huawei places the performance of the Kirin 990 5G at 2.5 times that of the Kirin 980, and an amount equivalent to that of the Snapdragon S855 and the Exynos 9825, a little less than 2x compared to the Apple A12 . Energy efficiency is also improved by similar amounts. All this consisted in comparing the inference scores to the FP16 and INT8 quantifications.

Kirin 990 and Kirin 990 5G: deadlines

These new chips are expected to be widely distributed on the flagship Huawei and Honor brands from here until 2019 and until 2020. Huawei organized a press conference in Munich on the 19th.^th from September, where we are waiting for the announcement of flagship phones Huawei Mate 30 and Mate 30 Pro – and probably a 5G model also, which could perhaps be only the 5G Pro.

We were told that the Kirin 990 4G chipset is ready and available. Due to other factors, probably related to market segmentation and strategy, Kirin 990 5G devices might be a little later.